Research On Person Re-identification Technology Based On Part Features

The person re-identification task is to continually track a person that appeared in non-overlapping cameras at distinct times.It has attracted more and more researchers to study on it due to its application in the security and surveillance systems,especially in the crowed places such as airports and subway stations.The research methods of person re-identification can be divided into two categories:supervised methods and unsupervised methods.The unsupervised methods consist of traditional machine learning methods and deep learning methods.In this paper,we use the deep learning as the tool to study this task with local characteristics of the human body.In the task of video-based person re-identification,most deep learning-based methods lay emphasis on extracting the feature of full body for creating sequence-level representation,in which the feature extracted by the methods are the representation of the full body.For clearly indicating the feature of person,we propose a novel Spatial and Temporal Features Mixture Model(STFMM),in which the model extracts the features of different parts of human body and then utilizes those features for ereating the person representation.The STFMM model fist horizontally splits the original video sequence into N parts video sequences.And then we use the deep learning networks to extract the feature information of each part video sequence.The feature of each part is integrated in order to achieve more expressive representation for each person.Experiments conducted on the iLIDS-VID and PRID-2011 datasets show that our approach achieves a rank-1 CMC accuracy of 73.6%and 81.3%respectively,which exceeds the jointly Attentive Spatial-Temporal Pooling Networks(ASTPN)by 11.3%and 4.0%.And the results of cross-dataset testing show that our model achieves a rank-1 CMC accuracy of 47.8%on the PRID-2011 dataset,which exceeds the most recent method ASTPN by 17.1%.The capsule network is a new type of deep learning network,which has a good performance in the image classification.We apply the capsule network to the person re-identification task for the first time.And we propose two models:CapsReld model and CapsPartReld model.The CapsReld model consist of the capsule network and the Siamese network,in which the capsule network is integrated into the subnet of the Siamese network.Combining the CapsReld model with the human body parts,we propose the CapsPartReld model in which the capsule network is integrated into the subnet of the developed Siamese network.We do a lot of experiments on the datasets VIPeR,CUHK01,GRID,PRID450S and CUHK03.From the results we can find that the re-identification accuracy of the CapsPartReld model exceeds the CapsReld model 3.7%,0.4%,1.3%,3.1%and 5.5%on CMC rank-1 respectively.